CN109508444B - Fast Tracking Method for Interactive Multimodal Generalized Label Multi-Bernoulli under Interval Measurement - Google Patents

Abstract

The invention discloses a fast tracking method for interactive multi-mode generalized label multi-Bernoulli under interval measurement, which combines the interactive multi-mode method and fast algorithm idea. First, under the framework of generalized label multi-Bernoulli filtering, For the target sampling particle prediction stage, combined with the interval measurement generalized likelihood function, the transition prediction of all particles to different models is realized, and then the model interaction is performed on the particles by calculating the model weight probability, and then the particles after model interaction are updated through the GLMB filter update equation. to update. On this basis, combined with the fast implementation method, the prediction and the update are combined, and only one truncation process is needed for each iteration, which reduces the calculation amount of the algorithm, and finally solves the detection and tracking problem of maneuvering multi-targets.

Description

Multi-Bernoulli fast-tracking method for interactive multimodal generalized labeling under interval measurement

technical field

The invention relates to the technical field of target tracking, in particular to a multi-Bernoulli fast tracking method for interactive multi-mode generalized labels under interval measurement.

Background technique

The motion of a non-maneuvering target can be described by a fixed model, but to describe the motion of a maneuvering target, it may be necessary to combine motion models with different maneuvering characteristics. As the maneuvering target tracking technology has received more and more attention, the requirements for the maneuvering target tracking technology are also getting higher and higher. Multi-maneuvering target tracking has become an extremely difficult problem in the field of target tracking.

In his paper "Target Tracking Algorithm of Box Particle Generalized Label Dobernoulli Filter" (Journal of Xi'an Jiaotong University, 2017, 51(10): 107-112.), Ji Hongbing proposed a box particle generalized label Dober Effort tracking algorithm. The algorithm uses the box particle filter algorithm to approximate the probability density of a single target state, that is, uses a set of uniform distributions with weights to fit the probability density of a single target state; Forecasting and updating, the position and velocity of a single target are estimated from the updated multi-target state probability density, and the track tracking can be realized because the labels of the single targets are different from each other. The disadvantage of this algorithm is that it cannot effectively track strong maneuvering targets.

In their paper "A Generalized Labeled Multi-Bernoulli Filter for Maneuvering Targets" (19th International Conference on Information Fusion), Vo et al. combined the concept of interactive multimodality with the theory of labeled multi-Bernoulli random finite sets (RFS) , a maneuvering target tracking algorithm based on Generalized Labeled Multi-Bernoulli Particle Filter (GLMB) is proposed, and the implementation form of Gaussian Mixture (GM) is given. The algorithm can track the tracks of different targets while estimating the states of multiple maneuvering targets. The disadvantage of this algorithm is that GLMB filter implementation faces huge computational complexity due to the combinatorial explosion problem in strong clutter and multi-object environment.

Vo et al. proposed a Gibbs sampling-based GLMB filter density truncation algorithm, which integrates prediction and update into one step, enables efficient implementation of GLMB filter. The disadvantage of this algorithm is that in many practical applications, standard measurement models are not sufficient. Although the sensor detection report is a point measurement value, the actual measurement will be affected by the unknown boundary error distribution, so that this non-standard measurement needs to be performed in the form of interval measurement.

Contents of the invention

What the present invention aims to solve is the problem that the target loses tracking when the target has a large maneuver under the interval measurement in the existing multi-maneuvering target tracking method, and provides an interactive multi-mode generalized tag multi-Bernoulli fast tracking method under the interval measurement. The tracking method combines interactive multi-mode methods and fast algorithm ideas on the basis of generalized label multi-Bernoulli filtering, and adopts sequential Monte Carlo (SMC) algorithm to achieve precise tracking of maneuvering multi-targets.

The basic idea of the present invention is: combine the interactive multi-mode method with the fast algorithm idea, firstly, under the framework of the generalized label multi-Bernoulli filter, aiming at the target sampling particle prediction stage, combined with interval measurement of the generalized likelihood function, to realize All particles are predicted for the transfer of different models, and then the model interaction is performed on the particles by calculating the model weight probability, and then the particles after the model interaction are updated through the GLMB filter update equation. On this basis, combined with the rapid implementation method, the prediction and update are combined, and only one truncation process is required for each iteration, which reduces the calculation amount of the algorithm, and finally solves the problem of maneuvering multi-target detection and tracking.

In order to solve the above problems, the present invention is achieved through the following technical solutions:

The multi-Bernoulli fast-tracking method for interactive multimodal generalized labeling under interval measurement includes the following steps:

Step 1. According to the target motion scene, set the state parameters of the target particles at the initial moment, and use the set state parameters as the initial distribution of the target particles, sample a fixed number of target particles with the initial distribution, and use the label multi-Bernoulli random The parameter set of the set is expressed in the form of a set, and the posterior distribution of the random set of the multi-Bernoulli random set of the target sampling particle label at the initial moment is obtained;

Step 2. Using the posterior distribution of the multi-Bernoulli random set of the target sampling particle label at the previous moment and the interval measurement data at the previous moment, use the interactive multi-mode method to predict the target sampling particle, and obtain the predicted The posterior distribution of a multi-Bernoulli random set of sampled particle labels for the target;

Step 3. Use the interval measurement data at the current moment to calculate the generalized likelihood function of each target sampling particle, and according to the generalized likelihood function, use the generalized label multiple Bernoulli filter to predict the target sampling particle label multiple at the current moment. The posterior distribution of the Bernoulli random set is updated to obtain the posterior distribution of the multi-Bernoulli random set of target sampling particle labels updated at the current moment;

Step 4. From the posterior distribution of the multi-Bernoulli random set of the target sampling particle label updated at the current moment obtained in step 4, select the target sampling particle label Do-Bernoulli updated at the current moment whose label weight is greater than a given threshold The posterior distribution of the random set is used as the posterior distribution of the multi-Bernoulli random set of the target sampling particle label truncated at the current moment;

Step 5. Calculate the potential of the posterior distribution of the multi-Bernoulli random set of target sampling particle labels truncated at the current moment obtained in step 5, and find out the index N corresponding to the maximum potential; the current From the posterior distribution of multi-Bernoulli random sets of target sampling particle labels truncated at time, select N-1 posterior distributions of multi-Bernoulli random sets of target sampling particle labels truncated at the current moment with larger label weights, as the current Posterior distribution of multi-Bernoulli random set of target sampled particle labels at time;

Step 6. Calculate the weighted sum of the posterior distribution of the final multi-Bernoulli random set of target sampling particle labels at the current moment, and use the weighted sum result as the estimated target state at the current moment;

Step 7. Judging whether all moments have been processed: if yes, output the target state estimated at the current moment; otherwise, execute step 2 to process the next moment.

Compared with prior art, the present invention has following characteristics:

1. Embed the multi-model concept and fast algorithm idea into the generalized label multi-Bernoulli algorithm, and adopt the particle filter method, which can not only accurately track the track of the target and estimate the number of targets, but also greatly reduce the calculation amount of the algorithm. It is smaller than the interactive multi-mode generalized label multi-Bernoulli algorithm, which improves the operation performance.

2. Combining interactive multi-mode, aiming at the target sampling particle prediction stage, it realizes the transfer prediction of all particles to different models, and then performs model interaction on the particles by calculating the model weight probability, realizing the detection and tracking of maneuvering targets, overcoming the current situation There is a problem of loss of tracking with technology to track maneuvering targets.

3. Combining the idea of fast algorithm, the target measurement model does not use the traditional sensor measurement model, the sensor output interval measurement, interval measurement to make up for any point measurement defects, fully considering factors such as the mobility instability of maneuvering targets , the fast algorithm integrates prediction and update into one step, and only needs one truncation process for each iteration, which greatly reduces the calculation amount of the algorithm, and effectively realizes the detection and tracking of maneuvering multi-targets under interval measurement.

Description of drawings

Figure 1 is a flow chart of the multi-Bernoulli fast-tracking method for interactive multimodal generalized labeling under interval measurement.

Fig. 2 is the emulation diagram of the present embodiment, wherein (a) is target motion locus figure, (b) is target track tracking result figure, (c) is target number tracking effect figure, (d) is target number tracking error figure , (e) is the target OSPA distance tracking error map, (f) is the target OSPA position tracking error map, (g) is the target OSPA potential tracking error map.

Detailed ways

In order to make the object, technical solution and advantages of the present invention clearer, the present invention will be further described in detail below in combination with specific examples and with reference to the accompanying drawings.

Aiming at the problems of interval measurement uncertainty and target mobility uncertainty, the present invention uses an interactive multi-model method to process the sampling particles of each target state in the filter on the basis of the generalized label multi-Bernoulli filter. Prediction, and then update the prediction particles by introducing the generalized likelihood function combined with the GLMB filter update strategy. On this basis, combined with a fast implementation method, combining prediction and update, only one truncation process is required for each iteration, which reduces the calculation amount of the algorithm. The present invention can effectively detect and track maneuvering multi-targets in the interval measurement environment , and the estimation of the target state and number is more accurate.

A fast-tracking method for multi-Bernoulli multimodal generalized labeling under interval measurement, the specific steps of which are as follows:

Step 1, initialize the target state.

According to the target motion scene, set the GLMB trajectory table at the initial moment, including target particles (including persistent living particles and newborn particles) and state parameters, that is, target position, velocity, weight, model probability, trajectory label and trajectory association history. Initialize GLMB component hypothesis weights, hypothesis labels, hypothesis potentials, and potential distributions. Use the parameters set above as the initial distribution of the target, use a Gaussian distribution to sample a fixed number of initial surviving particles and newborn particles, and express it in the form of a parameter set with a labeled multi-Bernoulli random set.

The initial particle sampling process is as follows:

为状态x₀对应的标签，

为初始时刻的标签集合。Among them, P ₀ is the covariance of the target state at the initial moment,

is the label corresponding to state x ₀ ,

is the set of labels at the initial moment.

Let the initial moment k=0, the initial distribution of the target is represented by the parameter set of the Gaussian particle label multi-Bernoulli random set as follows:

表示初始时刻GLMB轨迹表假设标签的假设权重；

表示初始时刻第j个目标采样粒子(标签为l)的模型权概率，

p为模型个数；

表示初始时刻第j个目标采样粒子的状态，

x₀表示初始时刻目标的横坐标，

表示初始时刻目标的水平速度，y₀表示初始时刻目标的纵坐标，

表示初始时刻目标的垂直速度，T表示转置操作；

表示初始时刻第j个目标采样粒子状态对应的状态权重；

表示初始时刻的目标采样粒子数目。where H is the indicator space, h is the indicator,

Indicates the assumed weight of the assumed label of the GLMB track table at the initial moment;

Indicates the model weight probability of the jth target sampling particle (labeled as l) at the initial moment,

p is the number of models;

Indicates the state of the jth target sampled particle at the initial moment,

x ₀ represents the abscissa of the target at the initial moment,

represents the horizontal velocity of the target at the initial moment, y ₀ represents the vertical coordinate of the target at the initial moment,

Indicates the vertical velocity of the target at the initial moment, and T indicates the transposition operation;

Indicates the state weight corresponding to the state of the jth target sampled particle at the initial moment;

Indicates the number of target sampling particles at the initial moment.

Step 2, predict and update the target state.

表示，其中(x_k,y_k)、

分别表示k时刻目标的位置、速度。目标为机动目标时，其运动模型随时间发生变化，运动方程为：The state of the target at time k can be represented by a 4-dimensional vector

Indicates that (x _k ,y _k ),

Respectively represent the position and velocity of the target at time k. When the target is a maneuvering target, its motion model changes with time, and the motion equation is:

x _k ＝f _k-1 (x _k-1 ,s _k )+v _k-1 (s _k )

Among them, f _k-1 represents the state transition equation of the target at time k, s _k represents the model variable at time k, and v _k-1 represents the state noise.

Assume that the posterior distribution of the multi-Bernoulli random set of target sampling particle labels at time k-1 is:

Then the multi-Bernoulli random set of persistent surviving particle labels is expressed as:

则k时刻预测的目标采样粒子标签多伯努利随机集为：The multi-Bernoulli random set of newborn particle labels at time k is

Then the multi-Bernoulli random set of target sampling particle labels predicted at time k is:

表示从k-1时刻到k时刻持续存活粒子在模型s下的状态预测，

表示k时刻新生目标的采样粒子在模型s下的状态预测，

表示k时刻持续存活粒子在模型s下的模型权概率，

表示新生目标的采样粒子k时刻在模型s下的模型权概率，

表示从k-1时刻到k时刻持续存活粒子在模型s下的权重预测，

表示k时刻新生目标的采样粒子在模型s下的粒子权重预测。

Indicates the state prediction of persistently surviving particles under model s from time k-1 to time k,

Represents the state prediction of the sampled particles of the newborn target at time k under the model s,

Indicates the model weight probability of persistently surviving particles at time k under model s,

Indicates the model weight probability of the sampling particle k of the newborn target under the model s,

Indicates the weight prediction of persistently surviving particles under model s from time k-1 to time k,

Represents the particle weight prediction of the sampled particles of the newborn target at time k under the model s.

The specific prediction method can be completed by the following steps.

Step 2.1: Resampling the multi-Bernoulli random set posterior distribution of target sampling particle labels updated at time k-1 to obtain the sampling samples of surviving particles at time k-1 as follows:

个新生粒子。Step 2.2, according to the Gaussian distribution, set 4 kinds of birth particle components, and sample them together

newborn particles.

Among them, N(m _i , P), i=1, 2, 3, 4 is the newborn density of the target at time k, and the specific process is expressed as follows:

个新生粒子。According to the mean value and variance of each birth particle component, particles are uniformly generated around, and the 4 birth particle components are co-generated

newborn particles.

结合交互式多模方法对存活粒子进行预测，具体方式表示如下：Step 2.3, resampling the surviving particle samples at time k-1, namely

Combining the interactive multi-mode method to predict the surviving particles, the specific method is expressed as follows:

Among them, F _s,k is the state transition equation corresponding to model s, s=1,...,p, p is the total number of models, and ν _k is the state noise.

Model weight probability calculation:

是预测模型概率，

T_r:,s为模型转移概率矩阵的第h列,

是对应模型s的预测粒子区间量测广义似然函数。in,

is the predictive model probability,

T _{r:, s} is the hth column of the model transition probability matrix,

is the predicted particle interval measurement generalized likelihood function corresponding to model s.

和模型权概率

可以得到交互多模型混合粒子，其中

具体计算如下：Particles predicted by each model

and model weight probabilities

Interactive multi-model hybrid particles can be obtained, where

The specific calculation is as follows:

In step 2.4, calculate the prediction status and weights of the survival target and the birth target:

in

Step 2.5, merge the weighted particles with labels:

Step 2.6, update the target state.

Assuming that the predicted target sampling particles at time k are expressed as:

Then the updated target sampling particle posterior distribution is:

表示从k-1时刻到k时刻第j个目标采样粒子状态预测，

表示k时刻第j个目标采样粒子状态的更新；

表示从k-1时刻到k时刻第j个目标采样粒子的权重的预测，

表示k时刻第j个目标采样粒子的权重的更新；

表示k时刻预测的目标采样粒子数目，

表k时刻的目标采样粒子数目。in,

Indicates the state prediction of the jth target sampled particle from time k-1 to time k,

Indicates the update of the state of the jth target sampled particle at time k;

Represents the prediction of the weight of the jth target sample particle from time k-1 to time k,

Represents the update of the weight of the jth target sampling particle at time k;

Indicates the number of target sampling particles predicted at time k,

Table k is the number of target sampling particles at time k.

Using the generalized likelihood function of the target random set at time k, update the multi-Bernoulli random set of target sampling particle labels predicted at time k, and obtain the posterior distribution of the multi-Bernoulli random set of target sampling particle labels at k time.

The specific update method can be accomplished by the following steps.

Step 2.6.1, using the interval measurement data at the current moment, calculate the generalized likelihood function corresponding to each predicted target sampling particle, assuming that, considering that on _Xk , the detection is independent, and the clutter has nothing to do with the detection, The specific calculation of multi-objective likelihood is as follows:

是一个函数，使得当i＝j时满足θ_k(i)＝θ_k(j)＞0，λ为杂波均值，本方法取λ＝10。其中in

is a function that satisfies θ _k (i)=θ _k (j)>0 when i=j, λ is the mean value of clutter, and this method takes λ=10. in

The single objective generalized likelihood function is

z＝[z_l,z_r]。where N(y;μ,P) denotes a Gaussian probability density function with mean μ and covariance P. Σ represents the covariance of the measurement noise v _k , that is, p _v (v) = N (v; 0, Σ). also

z=[z _l , z _r ].

Step 2.6.2, according to the interval measurement likelihood function, update and calculate the weight of each predicted target particle at time k:

为每个预测目标粒子的预测权重，其中M_s,k表示在k时刻模型变量的一个集合，M_s,k＝{s₁,…,s_k}。in,

is the prediction weight of each prediction target particle, where M _s,k represents a set of model variables at time k, M _s,k ={s ₁ ,…,s _k }.

Step 3, component truncation.

From the posterior distribution of the multi-Bernoulli random set of target sampling particle labels updated at the current moment, select the posterior distribution of the multi-Bernoulli random set of target sampling particle labels updated at the current moment whose label weight is greater than a given threshold, Posterior distribution for a Multi-Bernoulli random set of particle labels sampled as targets truncated at the current moment.

Usually, in the GLMB algorithm, the number of hypothetical trajectories at the previous moment will be truncated twice at the next moment, and the prediction and update of the hypothetical trajectories respectively require truncation operations. In order to improve the operation efficiency of the algorithm, the algorithm of the present invention integrates the prediction and updating steps, and only needs one truncation, which greatly improves the running speed of the algorithm.

Assume that the posterior distribution of the multi-Bernoulli random set of target sampling particle labels at time k-1 is:

Then the posterior distribution of the multi-Bernoulli random set of target sampling particle labels at time k is:

Where H is the index space, h is the index, from the total number of components at time k-1 is H _k-1 to the total number of components at time k is H _k , only one truncation operation is needed, which greatly reduces the computational complexity of the algorithm of the present invention.

Step 4, state estimation.

We use a suboptimal version of marginal multi-objective estimation, the posterior estimation of maximum potential. The average estimate of the multi-target state depends on the estimated potential, and the potential distribution is calculated for the updated target particles.

Calculate the potential of the posterior distribution of the multi-Bernoulli random set of target sampling particle labels truncated at the current moment, and find out the index N _k +1 corresponding to the maximum potential; From the posterior distribution of the random set, select the posterior distribution of N _k truncated target sampling particle labels at the current moment with a large label weight, and use it as the final target sampling particle label at the current moment. The posterior distribution for a random set.

According to the posterior distribution of the multi-Bernoulli random set of target sampling particle labels at the current moment, the weighted sum of the target sampling particle states is calculated by using the weighting method, which is regarded as the real target state at the current moment. If it is estimated at this time that the target number corresponding to the target posterior maximum potential is N _k . The specific calculation of state estimation is as follows:

为后验最大势对应假设轨迹的目标采样粒子的更新粒子，

表示对应更新粒子的更新权重。x_k,i为当前时刻第i个目标的状态估计。in,

is the update particle of the target sampling particle for the posterior maximum potential corresponding to the hypothesized trajectory,

Indicates the update weight of the corresponding update particle. x _k,i is the state estimation of the i-th target at the current moment.

Step 5, judge whether all time has been processed, if so, execute step 6, otherwise, execute step 2, process the next time.

Step 6, end.

Based on the generalized label multi-Bernoulli filter, the present invention uses an interactive multi-model method to predict the sampling particles of each target state in the filter, and then substitutes the predicted particles into the GLMB algorithm to perform target existence probability and distribution An updated estimate of the density. The invention can effectively detect and track maneuvering multi-targets under the condition of interval measurement environment, and realize the estimation of the status and number of the targets.

The effect of the present invention will be further described in conjunction with the simulation diagram of accompanying drawing 2 below.

Simulation conditions: the present invention uses MATLAB R2014a software to complete the simulation on a computer with an Intel(R) Core(TM) i3-2370M CPU@2.40GHz processor.

Simulation scene setting: In order to verify that the rapid implementation algorithm of an interactive multimode generalized marker multi-Bernoulli filter under interval measurement proposed by the present invention can accurately detect and track weak and small maneuvering targets, the simulation experiment scene of the present invention is [ In the two-dimensional space of -2000,2000]×[0,2000]m ² , the whole simulation process lasts for 100 seconds, and there are five targets in total. The birth, death time and motion state of each target are listed in Table 1.

Table 1

Target uniform linear motion left turn movement right turn movement 1 1～20s 21～65s 66～100s 2 10～25s 26～70s 71～100s 3 15～35s 36～65s 66～100s 4 20～35s 36～57s 58～80s 5 60～70s 71～95s 96～100s

The target state equation is:

表示运动模型s对应的状态转移矩阵，s＝1,2,3.

均为零均值高斯白噪声。定义

为匀速直线运动模型，

为左转弯运动，

为右转弯运动。in

Indicates the state transition matrix corresponding to the motion model s, s=1,2,3.

Both are zero-mean Gaussian white noise. definition

is a uniform linear motion model,

for a left turn movement,

Movement for a right turn.

为表示协方差矩阵为Q_j的零均值高斯白噪声，

T表示采样周期。其中σ_v＝4m/s²，ω ⁽²⁾ = -0.04rad/s, ω ⁽³⁾ = 0.1rad/s.

To represent zero-mean Gaussian white noise with covariance matrix Q _j ,

T represents the sampling period. where _σv = 4m/s ² ,

The measurement equation is:

表示量测似然函数，(x_k,y_k)表示目标的位置。其中ω_k为零均值且协方差矩阵为

的高斯白噪声，其中σ_r＝2.5m和σ_θ＝0.25°。这种传感器提供区间量测，具有区间长度为Δ＝[Δ_r,Δ_θ]^T，其中Δ_r＝50m和Δ_θ＝4°分别是范围和方位角的区间长度。in,

represents the measurement likelihood function, and (x _k , y _k ) represents the position of the target. where ω _k is zero mean and the covariance matrix is

Gaussian white noise of , where σ _r =2.5m and σ _θ =0.25°. This sensor provides interval measurements with interval length Δ = [Δ _r ,Δ _θ ] ^T , where Δ _r =50m and Δ _θ =4° are the interval lengths for range and azimuth, respectively.

The exact positions of five targets are given when the generalized label multi-Bernoulli filter is initialized, the initial state of the target is x ₁ = [1000,-10,1500,-10], x ₂ =[-250,20,1000,3 ], x ₃ =[-1500,11,250,10], x ₄ =[250,11,750,5], x ₅ =[1000,-50,1500,0]. The relevant simulation parameters are set as: p _D,k =0.98 and p _S,k =0.99, and the mean value of clutter λ=10. target initial model weight probability

In order to prove the simulation effect, the truncation threshold H _th =1×10 ^-15 , under the condition of clutter rate λ=10, the number of surviving and newly born particles is 100 particles and 100 Monte Carlo experiments are carried out. Calculate the OSPA (Optimal Sub-Pattern Assignment) distance of the target, and set the OSPA parameters as c=100m, p=2. Figure 2(a) and Figure 2(b) respectively show the motion state and track tracking results of the target when the clutter rate is λ=10. The circle in the trajectory of Figure 2(a) represents the initial position of the target, and the triangle represents the final position of the target. Fig. 2(b) is the track tracking result of the target; as can be seen from Fig. 2(b), the method of the present invention can deal with the problem of target maneuvering, and can also track the exact position of the target when the target turns. Figure 2(c) and Figure 2(d) are respectively the tracking effect diagram and the potential error diagram of the target potential; from Figure 2(c) and Figure 2(d), it can be seen that, without considering individual moments, the method of the present invention It shows a relative advantage, its estimation of the target potential is relatively more accurate, and the estimation error is relatively stable and smaller than the IMM-GLMB algorithm (interactive multi-mode generalized label multi-Bernoulli filter under interval measurement) and IMM-LMB Algorithm (interactive multimodal label multi-Bernoulli filter under interval measurement). Fig. 2(e), Fig. 2(f) and Fig. 2(g) respectively show the target OSPA distance tracking error map, OSPA position tracking error map and OSPA potential tracking error map when the number of surviving and newly born particles is 100; 2(e) shows that when the number of targets increases, the method of the present invention is relatively stable to the target distance tracking error and is smaller than the IMM-GLMB algorithm and the IMM-LMB algorithm; it can be seen from Figure 2(f) that the method of the present invention is When estimating the target position, the tracking performance is further improved, reducing the estimation deviation of the target maneuvering moment, and the filtering performance is more stable than the IMM-GLMB algorithm and the IMM-LMB algorithm; Fig. 2(g) It can be seen that the tracking performance of the method of the present invention is obviously better than that of the IMM-GLMB algorithm and the IMM-LMB algorithm when estimating the number of targets, and it can also track the number of targets well when the number of targets increases.

In summary, from the analysis of the simulation effect diagram, it can be seen that a fast tracking method for the interactive multimode generalized label multi-Bernoulli filter under the interval measurement proposed by the present invention realizes the mobile multi-mode filter under the interval measurement. Target detection and tracking. The target tracking accuracy is high, and the tracking performance is good, and its performance is relatively better than the interactive multi-mode generalized label multi-Bernoulli filter, the interactive multi-mode label multi-Bernoulli filter, the interactive multi-mode probability hypothesis density filter and the interactive The multi-mode potential probability assumption density filter, the method of the invention has obvious advantages in filtering performance when dealing with the problem of maneuvering target tracking, and has more advantages in practical engineering applications.

It should be noted that although the above-mentioned embodiments of the present invention are illustrative, they are not intended to limit the present invention, so the present invention is not limited to the above specific implementation manners. Without departing from the principles of the present invention, all other implementations obtained by those skilled in the art under the inspiration of the present invention are deemed to be within the protection of the present invention.

Claims (1)

1. The fast-tracking method of interactive multimode generalized label Multi-Bernoulli under interval measurement, it is characterized in that, comprises steps as follows: Step 1. According to the target motion scene, set the state parameters of the target particles at the initial moment, and use the set state parameters as the initial distribution of the target particles, sample a fixed number of target particles with the initial distribution, and use the label multi-Bernoulli random The parameter set of the set is expressed in the form of a parameter set, and the posterior distribution of the multi-Bernoulli random set of the target sampling particle label at the initial moment is obtained as:

where H is the index space, h is the index,

Represents the hypothesis weight of the hypothetical label of the generalized label multi-Bernoulli particle filter trajectory table at the initial moment;

Indicates the state of the jth target sampled particle at the initial moment,

x ₀ represents the abscissa of the target at the initial moment,

represents the horizontal velocity of the target at the initial moment, y ₀ represents the vertical coordinate of the target at the initial moment,

Indicates the vertical velocity of the target at the initial moment, and T indicates the transposition operation;

Indicates the model weight probability of the jth target sampling particle at the initial moment,

p is the number of models;

Indicates the state weight corresponding to the state of the jth target sampled particle at the initial moment;

Indicates the number of target sampling particles at the initial moment; l is the label; Step 2. Using the posterior distribution of the multi-Bernoulli random set of the target sampling particle label at the previous moment and the interval measurement data at the previous moment, use the interactive multi-mode method to predict the target sampling particle, and obtain the predicted The posterior distribution of the multi-Bernoulli random set of target sampled particle labels is:

In the formula,

Indicates the state prediction of persistently surviving particles under model s from time k-1 to time k,

Indicates the model weight probability of persistently surviving particles at time k under model s,

Indicates the weight prediction of persistently surviving particles under model s from time k-1 to time k,

Indicates the number of persistent living particles;

Represents the state prediction of the sampled particles of the newborn target at time k under the model s,

Indicates the model weight probability of the sampling particle k of the newborn target under the model s,

Represents the particle weight prediction of the sampled particles of the newborn target at time k under the model s;

Indicates the number of newborn particles; l is the label; The specific prediction method can be completed by the following steps: Step 2.1. Perform resampling on the multi-Bernoulli random set posterior distribution of the target sampling particle label updated at time k-1 to obtain a sampling sample of surviving particles at time k-1; Step 2.2. According to the Gaussian distribution, set 4 kinds of birth particle components and sample them together

a newborn particle; Step 2.3. Combine the sampling sample of surviving particles obtained by resampling at time k-1 with the interactive multi-model method to predict surviving particles, and the particles predicted by each model

and model weight probabilities

To get interactive multi-model hybrid particles, the specific calculation is as follows:

in,

is the predictive model probability,

T _{r:, h} is the hth column of the model transition probability matrix,

is the predicted particle interval measurement generalized likelihood function corresponding to model s,

F _s,k is the state transition equation corresponding to model s, s=1,...,p, p is the total number of models, ν _k is the state noise;

Step 2.4, calculate the prediction status and weight of survival target and birth target:

in

Step 2.5. Merge labeled weighted particles:

Step 3. Use the interval measurement data at the current moment to calculate the generalized likelihood function of each target sampling particle, and according to the generalized likelihood function, use the generalized label multiple Bernoulli filter to predict the target sampling particle label multiple at the current moment. The posterior distribution of the Bernoulli random set is updated, and the posterior distribution of the target sampling particle label multi-Bernoulli random set updated at the current moment is:

In the formula,

Indicates the update of the state of the jth target sampled particle at time k,

Represents the update of the weight of the jth target sampling particle at time k;

Indicates the number of target sampling particles predicted at time k,

Indicates the number of target sampling particles at time k; The specific update method can be completed by the following steps: Step 3.1. Using the interval measurement data at the current moment, calculate the generalized likelihood function corresponding to each predicted target sampling particle. The multi-target likelihood is specifically calculated as follows:

Step 3.2, according to the interval measurement likelihood function, update and calculate the weight of each predicted target particle at time k:

In the formula,

is a function such that θ _k (i)=θ _k (j)>0 is satisfied when i=j, λ is the mean value of clutter,

M _s,k represents a set of model variables at time k, M _s,k ={s ₁ ,…,s _k }; Step 4. From the posterior distribution of the multi-Bernoulli random set of the target sampling particle label updated at the current moment obtained in step 4, select the target sampling particle label Do-Bernoulli updated at the current moment whose label weight is greater than a given threshold The posterior distribution of the random set, as the posterior distribution of the multi-Bernoulli random set of the target sampling particle label truncated at the current moment is:

In the formula, Hk is the total number of components of the index space H at time k, h is the index,;

Indicates the hypothetical weight of the hypothetical label of the generalized label multi-Bernoulli particle filter trajectory table at time k,

Indicates the state of the jth target sampled particle at time k,

Indicates the model weight probability of the jth target sampling particle at time k,

Indicates the state weight corresponding to the state of the jth target sampled particle at time k;

Indicates the number of target sampling particles at time k; l is the label; Step 5. Calculate the potential of the posterior distribution of the multi-Bernoulli random set of target sampling particle labels truncated at the current moment obtained in step 5, and find out the index N corresponding to the maximum potential; the current From the posterior distribution of multi-Bernoulli random sets of target sampling particle labels truncated at time, select N-1 posterior distributions of multi-Bernoulli random sets of target sampling particle labels truncated at the current moment with larger label weights, as the current Posterior distribution of multi-Bernoulli random set of target sampled particle labels at time; Step 6. Calculate the weighted sum of the posterior distribution of the final multi-Bernoulli random set of target sampling particle labels at the current moment, and use the weighted sum result as the estimated target state at the current moment; the specific calculation of the state estimation is as follows:

Among them, x _k,i is the state estimation of the i-th target at the current moment,

is the update particle of the target sampling particle for the posterior maximum potential corresponding to the hypothesized trajectory,

Indicates the update weight of the corresponding update particle;

Indicates the number of target sampling particles at time k; l is the label; N _k is the number of targets corresponding to the posterior maximum potential; Step 7. Judging whether all moments have been processed: if yes, output the target state estimated at the current moment; otherwise, execute step 2 to process the next moment.

Images